Case Study · AI Automation · Business Process Outsourcing
AI-Powered Process Automation for a Global BPO Company
Capturing how work really gets done, and automating it, for a global BPO provider
How XYZ turned 146 hours of recorded client work into step-by-step SOPs, process insights and reliable automations across 16 client engagements.
The result: 618 recordings and about 146 hours of real client work captured across 50 workflows and 16 client engagements, with 10 computer-use automations completing runs across 5 engagements.
The Results
What the AI process capture delivered
Figures from the live deployment, June to October 2026:
- 618
- recordings of real client work, about 146 hours in total, captured by analysts across the business units
- 50
- distinct workflows captured across 16 client engagements
- 233,000+
- individual steps captured, with nearly 219,000 screenshots and supporting files: the raw material for SOPs, training and automation
- 81%
- of recordings include the analyst's spoken explanation, preserving the judgment behind the clicks
- 29%
- of recordings (178) were made inside remote desktop sessions, where conventional recording tools capture nothing they can use
- 10
- automations with completed runs across 5 client engagements, including a read-only check of a retail client's HR and benefits systems that has completed 98 runs since July 2026
Process knowledge that used to live in people's heads is now recorded, searchable and shared. New SOPs are produced as a by-product of doing the work rather than written afterwards, team leads can see how work is really done across their clients, and automation now starts from evidence instead of discovery workshops.
The Customer
A global business process outsourcing provider
The customer is a global business process outsourcing (BPO) provider. Its delivery teams carry out high-volume back-office work for clients in North America, the UK and Europe, working directly inside each client's own systems.
The work is broad: claim follow-up, eligibility checks and denial management for healthcare providers; insurance operations; customs and shipping documentation for logistics companies; HR and benefits administration for a retailer; and programme administration for an energy-services client. The company is organised into eight business units, each running several client engagements with their own systems, rules and quality standards.
At a glance
- Customer
- A global business process outsourcing (BPO) provider
- Work covered
- Healthcare revenue cycle, insurance operations, logistics and customs documentation, retail HR and benefits, and energy-programme administration
- Users
- Analysts and team leads across eight business units
- Solution
- XYZ, Decision Foundry's platform for capturing, documenting and automating desktop work
- Deployment
- A Windows desktop app on each analyst's machine and a central hub in the customer's own cloud environment
The Problem
Most of the knowledge behind the work lived in analysts' heads
Most of the knowledge behind this work lived in analysts' heads. A process was typically taught over a call, written up once in a document that soon went out of date, and then passed on informally from one analyst to the next. This created three business problems:
- Slow and uneven onboarding. New analysts learned by shadowing experienced colleagues, so the quality of their work depended on who trained them.
- No clear picture of how work was really done. Leaders could not see how many different ways the same task was being done, where analysts were working around a system, or where errors crept in.
- Automation that was hard to start. The company wanted to automate repetitive work, but every automation effort began with long discovery sessions in which experts explained a process while someone took notes. Much of the work also ran in systems that standard automation tools could not reach.
The company needed a way to capture how work is actually done, across hundreds of processes and dozens of analysts, and to turn that into documentation and automation without taking experts away from client work.
Challenges Faced Building the AI Process Automation
Work happens inside client-controlled systems.
Analysts use web portals, legacy Windows applications and, very often, remote desktop sessions into the client's environment. Inside a remote session there is no web page or application structure to read, only an image of the screen, so most recording and robotic process automation (RPA) tools cannot see what the analyst is doing.
Sensitive data and strict client separation.
Much of the work involves patient, financial and employee data. Each client's recordings had to be visible only to that client's team, passwords could never be stored, and all data had to stay in the customer's own cloud.
Scale without disruption.
Dozens of analysts working different shifts needed to record their work on their own machines without IT tickets, complex set-up or any slowdown in their day-to-day work.
The “why” is invisible.
Clicks show what an analyst did, not why they did it: which status code to choose, when to escalate, what to check before submitting. This judgment is the most valuable part of a process and the part most often lost.
The same task is done many ways.
Before automating anything, the team needed to know which version of a process was the standard and which variations were mistakes, workarounds or legitimate exceptions.
Automation must be affordable and safe.
AI agents that interpret the screen at every step are costly to run and can behave unpredictably, while scripts recorded the traditional way break as soon as a screen changes. A wrong click in a client's billing system is not acceptable.
The AI Solution
XYZ: AI-assisted capture, documentation and automation of desktop work
Decision Foundry deployed XYZ, its AI-assisted platform for capturing, documenting and automating desktop work. XYZ is built on one idea: watch the work once, then reuse it as documentation, as process insight and as automation. Anthropic's Claude models help turn recordings into documentation and analysis; Decision Foundry's automation consultants build the automations from that captured knowledge.
01
One-click recording in any application
Analysts install a lightweight Windows app and press Record. XYZ captures every click, keystroke and screen across web portals, desktop applications and remote desktop sessions, and chooses the right capture method for each window automatically. Analysts talk through what they are doing as they work. XYZ transcribes the narration and links it to the matching steps, so the reasoning behind each decision is captured along with the actions. Password fields are masked at the moment of capture and are never stored.
02
A central hub, organised by client
Recordings upload automatically to a central hub hosted in the customer's own cloud. Work is organised by client, project and workflow, and access is scoped so that each team sees only its own clients. New analysts are set up from the customer's internal request portal, so they can start recording without manual set-up. Team leads can see at a glance which workflows are documented, which need more recordings and which are worth automating.
03
SOPs written automatically from real work
When a recording finishes, XYZ produces a standard operating procedure (SOP): numbered steps grouped into phases, screenshots with the clicked control highlighted, the analyst's own explanations, decision tables and note templates. The steps, their order and the controls used all come from what was actually recorded; AI writes the explanations around them, not the steps themselves. As a result, the document matches what happened on screen. SOPs can be shared with a link or exported. When many analysts record the same task, their recordings can also be compared side by side, as the clear win below shows.
04
Automations built by consultants from captured knowledge
For workflows worth automating, Decision Foundry's automation consultants build the automations. XYZ's MCP connection (Model Context Protocol) gives them everything XYZ captured, directly in their development environment: every recorded step and screen, the analysts' narration, the SOPs and the analysis of how the work varies. Consultants work from this evidence to build computer-use automations (CUA), which operate the same screens an analyst uses (web portals, desktop applications and remote desktop sessions) without needing APIs or changes to the client's systems.
Each automation runs as a fixed, repeatable sequence of steps, so it behaves the same way every time and stays inexpensive to run. Safety is built in. Before every action, XYZ checks that the screen matches the one in the recording. If it does not match, XYZ stops without clicking and hands the record to an analyst along with a screenshot and the context. Automations can also be limited to reading and reporting, so they never change client records. Every run, across every client, appears in a single oversight view.
See the AI in Action
1. Recording a process inside a remote desktop session.
An analyst records a claim follow-up that moves between a legacy billing application (running in a remote desktop session) and a payer's web portal. XYZ detects the remote session and switches to screen-and-text capture on its own. The recording panel shows the live transcript of the analyst's explanation, the number of steps captured so far, and confirmation that password fields are being masked.
2. A client workspace on the XYZ hub.
Each client has its own workspace. A team lead can see every workflow for the client, how many recordings and analysts are behind it, whether its SOP is ready, its estimated automation potential and whether it has already been automated. Only the people assigned to this client can open it.
3. An SOP generated from analysts' recordings.
XYZ turns recordings into a step-by-step SOP grouped into phases. Each step shows the screen with the relevant control highlighted, the analyst's own explanation taken from the narration, and a decision table for the judgment calls. This example was built from 18 recordings by four analysts.
4. Overseeing automation runs across clients.
Every automation run across all clients appears in one view. When a screen does not match what was recorded, the automation pauses before acting and passes the record to an analyst with the expected and live screens side by side. When a portal moves a control, XYZ repairs the automation, the run carries on, and the repair is logged for review.
A Clear Win
A custom AI process analysis for a customs brokerage
One of the customer's clients, a North American customs brokerage, wanted to see how its back-office work was really done and what it would take to automate it. Decision Foundry analysed 429 XYZ recordings across three of its workflows: customs document indexing (306 recordings), release processing (106) and contact-centre customer service (17). For each workflow, the client received a custom report covering:
- The happy path: the route a clean transaction takes, step by step, as analysts actually did it.
- The divergent paths: every way real work departed from it, how often and why, each linked to the recordings behind it.
- A coverage plan: how many divergent paths an automation must handle, on top of the happy path, to reach an 80% target.
- Recommendations: where to start, and the business rules no recording contains, which the client must settle first.
The findings reshaped the plan
Indexing proved to be two separate processes in two applications, so it needs two builds, not one. One lane runs clean 55% of the time and reaches 80% with six divergent paths; the other runs clean 35% of the time and needs seven, so the first lane is the natural place to start. Release processing runs clean only 10% of the time, with nearly half of its files ending on hold, so its hold rules need settling first. With this evidence, the customer and its client can shape an automation strategy tailored to the client: which lane to build first, which rules to fix and which scenarios to record next.
The AI Process Analysis Report
5. The indexing analysis delivered to the customs brokerage.
The main page gives the short version, the two lanes, the coverage plan against the 80% target, the happy path, the questions only the client can answer and a recommendation. Every number opens the recordings behind it.
About XYZ
XYZ is Decision Foundry's platform for capturing, documenting and automating desktop work. Analysts record their work once, in any web, desktop or remote desktop application, and XYZ turns it into SOPs, process insights and the evidence consultants use to build reliable automations.
Get In Touch